Skip to main content

Gemini 4 Argon: One Output, a Million Tokens

Google DeepMind has quietly dropped a bombshell: Gemini 4 Argon, a model that can produce a staggering 1 million tokens in a single output. That's not the context window—it's how much it can write in one go. Most top models cap out around 128,000 tokens, so Argon's leap is like going from a paragraph to a novel without taking a breath.

Why does that matter? Imagine migrating an entire codebase, drafting a lengthy technical report, or automating a multi-step process. With Argon, you don't have to chop the task into pieces. It can carry the thread from start to finish, keeping everything coherent. That's a game-changer for professionals who deal with sprawling, interconnected work.

Built for the Long Haul

Argon isn't for casual chitchat. Google is positioning it for "complex, long-term workflows"—think software engineering, law, finance, and cybersecurity defense. The benchmarks back it up: a record 77.9% on DeepSWE v1.1 for long-horizon software tasks, first place on the Vals Index (which weighs finance, programming, law, and taxation), and a top score of 51.3% on Zapier's AutomationBench. It also aced long-video understanding with 91.7% on LVBench, and can handle professional charts and cross-document actions.

But Google isn't flinging the doors open. Argon is being released in phases: first to trusted cybersecurity professionals, then to developers, enterprises, and subscribers. That cautious approach reflects the model's power—and the risks.

Security at the Core

Cybersecurity is the headline here. Argon is trained to autonomously find, verify, and fix critical vulnerabilities. For vetted defenders and Google's internal teams, there's even a version with "no network barrier." Security firm Wiz used it through the Scan for Good project to uncover a serious flaw in a global hospital's medical software—one that previous cutting-edge models had missed. And on CWE-bench v1, Argon ranked first for vulnerability repair with a 68% score.

Before wide release, Google is layering on four advanced defenses: preventing abuse, blocking prompt injection, monitoring misaligned reasoning chains, and reinforcing sandboxes. Thousands of Google employees are already putting Argon through its paces—debugging code, designing algorithms, and migrating massive codebases. In one case, quantum algorithms improved by 40% in minutes.

Pricing and Availability

For early adopters, Argon costs $2 per million input tokens and $10 per million output tokens. After the introductory period, those prices will double to $4 and $20, respectively. It's a premium tool for premium tasks.

So no, Argon isn't here to chat about the weather. It's here to tackle the hardest problems professionals face—and it's just getting started.

Key Points:

  • 1M token output limit—far beyond the typical 128K, enabling seamless long-form work.
  • Top benchmark scores in software engineering, automation, and long-video understanding.
  • Phased rollout starting with cybersecurity experts, then developers and enterprises.
  • Security-first design with autonomous vulnerability detection and multiple defense layers.
  • Introductory pricing at $2/$10 per million tokens, doubling later.